Learning a hierarchical compositional representation of multiple object classes

Abstract

Summary form only given. Visual categorization, recognition, and detection of objects has been an area of active research in the vision community for decades. Ultimately, the goal is to recognize and detect a large number of object classes in images within an acceptable time frame. This problem entangles three highly interconnected issues: the internal object representation which should expand sublinearly with the number of classes, means to learn the representation from a set of images, and an effective inference algorithm that matches the object representation against the representation produced from the scene. In the main part of the talk I will present our framework for learning a hierarchical compositional representation of multiple object classes. Learning is unsupervised, statistical, and is performed bottom-up. The approach takes simple contour fragments and learns their frequent spatial configurations which recursively combine into increasingly more complex and class-specific contour compositions.